AutoML-Med: Medical Data Pipeline Optimization

• AutoML-Med is an automated ML framework tailored to medical data, optimizing preprocessing and model selection to address missing values, imbalance, and high-dimensional features.
• The framework employs a staged workflow with Latin Hypercube Sampling to efficiently explore and refine preprocessing configurations based on their impact on predictive metrics.
• Empirical evaluations in Multiple Sclerosis and Type 2 Diabetes demonstrate enhanced sensitivity and balanced accuracy compared to traditional methods, despite some trade-offs in specificity.

AutoML-Med denotes automated machine-learning methods tailored to medical data and, in the most specific sense represented in the recent literature, refers to a framework for medical tabular data that automates preprocessing selection, model selection, hyperparameter tuning, and targeted refinement under conditions typical of healthcare datasets: missing values, class imbalance, heterogeneous feature types, and high dimensionality relative to sample size (Francia et al., 4 Aug 2025). Within the wider medical AI literature, the term also sits inside a broader AutoML program spanning clinical tabular prediction, medical imaging, radiomics, pathology, and increasingly agentic pipeline construction, all motivated by the mismatch between the complexity of medical data and the limited availability of ML engineering expertise in many clinical and translational settings (Jidney et al., 2023).

1. Problem formulation and domain rationale

AutoML-Med is designed around the observation that, in medical tabular problems, predictive performance depends not only on the final estimator but on the interaction between preprocessing and modeling. The framework explicitly targets missing values, class imbalance, heterogeneous feature types, high dimensionality with small sample size, redundant or irrelevant variables, noisy features, and inconsistent entries (Francia et al., 4 Aug 2025). This emphasis reflects a recurrent pattern in medical ML: data preparation is often at least as consequential as classifier choice.

Related studies reinforce that premise. In cardiovascular disease detection with AutoGluon, three preprocessing scenarios produced materially different outcomes, with AutoML test accuracy ranging from 87.41% to 92.3%; the best result came from manual transformation into binary values via intervaling and One-Hot Encoding, while built-in AutoML preprocessing was not optimal (Afanasieva et al., 2023). In administrative claims prediction over 12,425,832 insured individuals and 3,511 engineered features, three general-purpose AutoML frameworks improved over a random forest baseline, yet none consistently outperformed the others and all produced low AUCPR under disease prevalences from 0.053% to 0.63%, underscoring how extreme imbalance and feature noise remain difficult even for mature AutoML systems (Romero et al., 2021).

This body of evidence suggests that AutoML-Med should be understood less as a generic “one-click” optimizer than as a medical-data-specific search framework in which preprocessing is a first-class optimization target.

2. Architecture of the AutoML-Med framework

AutoML-Med implements a staged workflow organized as phases (a)–(e), with the pipeline space defined over five ordered preprocessing steps followed by model training and refinement (Francia et al., 4 Aug 2025). The ordered preprocessing steps are:

1. Missing data imputation
2. Class balancing
3. Feature engineering
4. Feature scaling
5. Feature selection

The framework is meant to minimize user intervention while preserving limited configurability. It can automate preprocessing selection, model selection, hyperparameter tuning, metric-based evaluation, and final refinement; at the same time, it allows users to choose which preprocessing methods to include, skip steps, select evaluated models, choose target metrics, run on a sampled subset for speed, and set a global seed for reproducibility (Francia et al., 4 Aug 2025).

Phase	Function
(a)	Latin Hypercube Sampling of preprocessing configurations
(b)	Application of sampled preprocessing pipelines
(c)	Model training and evaluation
(d)	PRCC-based sensitivity analysis
(e)	Fine-tuned grid search on influential preprocessing steps

A central design choice is that AutoML-Med does not attempt exhaustive optimization over the full Cartesian product of preprocessing methods. Instead, it first samples broadly, then trains predictive models on the sampled variants, and only afterward concentrates search effort on the preprocessing stages shown to exert the strongest influence on the chosen metric. This staged structure distinguishes it from fixed-pipeline medical AutoML use and from approaches that primarily optimize the estimator layer.

3. Search strategy, objective functions, and refinement

The initial exploration mechanism is Latin Hypercube Sampling (LHS). Because the preprocessing search space is the Cartesian product of candidate methods across the five ordered stages, exhaustive enumeration would be expensive. AutoML-Med therefore uses LHS as a stratified sampling strategy that covers the space more uniformly than random sampling. For a preprocessing step $s$ with $N_s$ possible methods, the algorithm samples

$x_s \in [0, N_s)$

and discretizes the sampled real value by taking its integer part, thereby selecting a method for that step (Francia et al., 4 Aug 2025). The result is a computationally manageable but broad subset of pipeline configurations.

For each sampled pipeline, one or more predictive models are trained. Hyperparameter tuning is performed through a random cross-validation search, and evaluation is parallelized to reduce runtime. Because the target setting is medical classification with imbalance, AutoML-Med emphasizes metrics beyond ordinary accuracy: Balanced Accuracy, F1-score, $F_\beta$-score with $\beta = 0.5$, and Matthews Correlation Coefficient (MCC) (Francia et al., 4 Aug 2025). Balanced accuracy is explicitly defined as

$$\text{Balanced Accuracy} = \frac{\text{Sensitivity} + \text{Specificity}}{2}.$$

After model selection, the framework applies Partial Rank Correlation Coefficient (PRCC) to quantify which preprocessing steps most influence the target metric. Let $Y$ denote the target metric and $\{X_1,\dots,X_k\}$ the preprocessing stages. For a given $X_i$, AutoML-Med regresses $X_i$ against the remaining inputs to obtain residual $N_s$0, regresses $N_s$1 against the same remaining inputs to obtain residual $N_s$2, and computes

$N_s$3

PRCC values near $N_s$4 indicate strong positive influence, values near $N_s$5 strong negative influence, and values near $N_s$6 little or no influence (Francia et al., 4 Aug 2025). The framework then refines the $N_s$7 most influential preprocessing stages, with $N_s$8 by default, through a grid search over $N_s$9 variations while holding the other stages fixed. In effect, AutoML-Med combines broad LHS exploration with local, sensitivity-guided exploitation.

4. Empirical evaluation on clinical risk prediction

AutoML-Med was evaluated in two clinical settings: Multiple Sclerosis (MS) risk prediction and Type 2 Diabetes risk prediction (Francia et al., 4 Aug 2025).

In the MS study, the dataset comprised 1,031 real patients, 33 features, and a positive/negative ratio of 0.27. The target was prediction of risk of reaching disability level 4 on the EDSS scale. Evaluation used 10 independent runs with varying seeds and a stratified 2/3 training, 1/3 testing split. Compared with Auto-sklearn, GAMA, and AutoBalance, AutoML-Med achieved the highest Balanced Accuracy at 0.8896 and the highest Sensitivity at 0.8539. Its Specificity was 0.9253, F1 macro was 0.8745, and MCC was 0.7558 (Francia et al., 4 Aug 2025).

In the Type 2 Diabetes study, the public BRFSS 2014 dataset began with 464,664 samples and 279 variables; after filtering, it contained 27 variables and final class counts of 76,156 negative and 14,532 positive, with a positive/negative ratio of 0.19. Using the same 2/3 training, 1/3 testing protocol and 10 independent runs, AutoML-Med was compared against Logistic Regression, Linear SVM, RBF SVM, and Neural Network models reported by Xie et al. AutoML-Med achieved Balanced Accuracy 0.7436 and Sensitivity 0.7968, both the best values in the comparison; its Specificity was 0.6904 and AUC was 0.7436 (Francia et al., 4 Aug 2025).

Clinical setting	Main result
Multiple Sclerosis	Balanced Accuracy 0.8896, Sensitivity 0.8539
Type 2 Diabetes	Balanced Accuracy 0.7436, Sensitivity 0.7968

These experiments illustrate the framework’s declared bias toward clinically consequential recall. In both settings, AutoML-Med improved sensitivity and balanced accuracy relative to alternatives, even when this entailed a reduction in specificity in the diabetes task. The authors explicitly interpret that trade-off as acceptable for identifying at-risk patients in medical settings (Francia et al., 4 Aug 2025).

5. Position within the broader medical AutoML landscape

AutoML-Med belongs to a heterogeneous medical AutoML ecosystem rather than an isolated technical line. A general survey of medical imaging AutoML characterizes the field as automating feature engineering, algorithm selection, hyperparameter optimization, and increasingly neural architecture search for modalities such as MRI, CT, and X-ray, while stressing persistent issues of privacy, heterogeneity, class imbalance, and interpretability (Jidney et al., 2023). The MedMNIST Classification Decathlon operationalized this agenda as a lightweight benchmark of 10 standardized $x_s \in [0, N_s)$0 datasets for medical image classification and showed that no single method generalized best across all tasks, although Google AutoML Vision was often strong (Yang et al., 2020).

Subsequent work diversified the technical repertoire. T-AutoML searched architecture, augmentation, and hyperparameters jointly for 3D lesion segmentation and reported a 0.7650 lesion Dice score on LiTS, slightly above nnU-Net at 0.7630 (Yang et al., 2021). AMDet, an evaluation of Microsoft Azure AutoML for mitotic cell detection in breast cancer histopathology, found that the best configuration used 256×256 unnormalized patches and achieved mAP = 0.629 at IoU 0.5, while larger patch sizes degraded performance (Williams et al., 2021). In radiomics, a comparative study across 10 datasets reported that Simplatab achieved the highest average test AUC at 81.81%, whereas LightAutoML reached competitive performance with approximately 6 minutes runtime and 78.74% mean AUC in six minutes in the abstract summary (Lozano-Montoya et al., 13 Jan 2026).

Another trajectory moves from AutoML search to autonomous pipeline generation. M$x_s \in [0, N_s)$1Builder introduced a medical-imaging workspace with four specialized agents and achieved a 94.29% success rate using Claude-3.7-Sonnet on M$x_s \in [0, N_s)$2Bench (Feng et al., 27 Feb 2025). AMID extended this trend with data-conditioned method planning and verification-guided optimization, producing valid accepted results for all 20 tasks in the ReX-MLE benchmark and outperforming the strongest listed agent baseline on 19 of 20 tasks (Liu et al., 12 Jul 2026). A plausible implication is that AutoML-Med, originally formulated for medical tabular data, now occupies one part of a larger transition from estimator-centric automation toward domain-aware, auditable, and modality-specific ML orchestration.

6. Limitations, misconceptions, and open directions

A common misconception is that medical AutoML is intrinsically plug-and-play. The literature does not support that view. AutoML-Med itself samples only a subset of the preprocessing space in its initial LHS stage, so the global optimum may be missed; its performance also depends on the supplied preprocessing-method lists, and specificity can decrease when sensitivity is prioritized (Francia et al., 4 Aug 2025). In the cardiovascular study, the best result did not come from the system’s built-in preprocessing but from a manually designed binary-feature pipeline, while the scenario with the highest training accuracy (95.65%) generalized worst on test (87.41%) (Afanasieva et al., 2023).

Another misconception is that ranking metrics alone determine clinical usefulness. The claims benchmark shows otherwise. Although AutoML tools improved ROC AUC over random forest, AUCPR remained low and no framework consistently dominated. More importantly, the paper demonstrated that threshold choice materially changes clinical utility: for lung cancer, a hypothetical $x_s \in [0, N_s)$3500 test cost favored the 90th percentile cutoff with FPR = 0.1 and TPR = 0.52 (Romero et al., 2021). This indicates that AutoML model selection and operating-point selection are distinct problems in healthcare.

Open technical gaps also remain modality-specific. In radiomics, current frameworks still show limited support for survival analysis, weak integration of harmonization, and inadequate treatment of feature reproducibility (Lozano-Montoya et al., 13 Jan 2026). In agentic medical imaging, AMID argues that optimization evidence is not trustworthy unless validation protocol, metric computation, and prediction artifacts are independently verified; reviewer-accepted evidence, rather than raw score maximization, becomes the basis for promotion and final selection (Liu et al., 12 Jul 2026). Taken together, these findings suggest that the next phase of AutoML-Med research is likely to emphasize auditable validation, stronger handling of imbalance and drift, richer preprocessing search, and domain-specific workflow control rather than reliance on default end-to-end automation alone.